technology

The “chain of thought” output simply gives you the “progress” and the specific path/approach the model has arrived at a particular answer - which is useful for tweaking and troubleshooting the parameters toward improving the accuracy and reducing hallucinations on a model, but it is not the same reasoning that could be given from a human mind.

The transformer architecture is really just a statistical model built to have very strong memory retention when it comes to making associations (in the case of LLMs, words). It fundamentally cannot think or reason. It takes a specific “statistical” path and arrives at an answer based on the associations it has been trained on, but you cannot make it think and reason the way we do, nor can it evaluate or verify the validity of a piece of information based on cognitive reasoning.

That's fair, but human oversight doesn't mean they'll necessarily catch biases in its output

sorry sweaty, ssds are problematic

Sufi whirling is the greatest expression of revolutionary spirit in all of time.

hey man come here i have something to show you

Future senator getting "show hog" to some question with 0.000001 probability: well, if the god-machine says so

Put it that way, even if one feeds it cia files to the hearts content, the weights of words which are needed to construct sentences is still sitting somewhere there. (also answering about real role of cia implies llm has any idea about reality, it will just bias answer in another direction, just as marxist analysis: it will just reproduce likeliest answer resembling marxist literature you fed to it, not "have analysis").

Benign application of llm is natural language processing into fixed functions on the back end (e.g. turn off the lights when it start raining or whatever, something which can be disassembled from millions of ways into same set of instructions, here its fuzziness is great)

Feed different data to the machine and you will get different results.

These things have already eaten all the data that there is, and I don't need to tell you that, but that data, as it has been produced almost solely under capitalism, is just crap.

We've been seeing great strides in reducing that power hunger recently, including by the LLM that's the subject of this post.

Due to the market economy in both the United State and China, further development of LLM efficiency is probably the worst thing that could possibly happen. Even if China did not want to subject LLMs to market forces, they are going to need to compete with the US. This is going further accelerate the climate disaster.

I should have been more precise, but this is all in the context of news about a cutting-edge LLM using a fraction of the cost of ChatGPT, and comments calling it all "reactionary autocorrect" and "literally reactionary by design".

I disagree that it's "reactionary by design". I agree that it's usage is 90% reactionary. Many companies are effectively trying to use it in a way that attempts to reinforce their deteriorating status quo. I work in software so I always see people calling this shit a magic wand to problems of the falling rate of profit and the falling rate of production. I'll give you an extrememly common example that i've seen across multiple companies an industries.

Problem: Modern companies do not want to be responsible for the development and education of their employees. They do not want to pay for the development of well functioning specialized tools for the problems their company faces. They see it as a money and time sink. This often presents itself as:

• missing, incomplete, incorrect documentation
• horrible time wasting meeting practices

I've seen the following be pitched as AI Bandaids:

Proposal: push all your documentation into a RAG LLM so that users simply ask the robot and get what they want

Reality: The robot hallucinates things that aren't there in technical processes. Attempts to get the robot to correct this involves the robot sticking to marketing style vagaries that aren't even grounded in the reality of how the company actually works (things as simple as the robot assuming how a process/team/division is organized rather than the reality). Attempts to simply use it as a semantic search index end up linking to the real documentation which is garbage to begin with and doesn't actually solve anyone's real problems.

Proposal: We have too many meetings and spend ~4 hours on zoom. Nobody remembers what happens in the meetings, nobody takes notes, it's almost like we didn't have them at all. We are simply not good at working meetings and it's just chat sessions where the topic is the project. We should use AI features to do AI summaries of our meetings.

Reality: The AI summaries cannot capture action items correctly if at all. The AI summaries are vague and mainly result in metadata rather than notes of important decisions and plans. We are still in meetings for 4 hours a day, but now we just copypasta useless AI summaries all over the place.

Don't even get me started on CoPilot and code generation garbage. Or making "developers productive". It all boils down to a million monkey problem.

These are very common scenarios that I've seen that ground the use of this technology in inherently reactionary patterns of social reproduction. By the way I do think DeepSeek and Duobao are an extremely important and necessary step because it destroys the status quo of Western AI development. AI in the West is made to be inefficient on purpose because it limits competition. The fact that you cannot run models locally due to their incredible size and compute demand is a vendor lock-in feature that ensures monetization channels for Western companies. The PayGo model bootstraps itself.